NEURAL ENCODING OF FELT AND IMAGINED TOUCH WITHIN HUMAN POSTERIOR PARIETAL CORTEX

In the human posterior parietal cortex (PPC), single units encode high-dimensional information with partially mixed representations that enable small populations of neurons to encode many variables relevant to movement planning, execution, cognition, and perception. Here we test whether a PPC neuronal population previously demonstrated to encode visual and motor information is similarly selective in the somatosensory domain. We recorded from 1423 neurons within the PPC of a human clinical trial participant during objective touch presentation and during tactile imagery. Neurons encoded experienced touch with bilateral receptive fields, organized by body part, and covered all tested regions. Tactile imagery evoked body part specific responses that shared a neural substrate with experienced touch. Our results are the first neuron level evidence of touch encoding in human PPC and its cognitive engagement during tactile imagery which may reflect semantic processing, sensory anticipation, and imagined touch.

In the human posterior parietal cortex (PPC), single units encode high-dimensional information with 28 partially mixed representations that enable small populations of neurons to encode many variables 29 relevant to movement planning, execution, cognition, and perception. Here we test whether a PPC 30 neuronal population previously demonstrated to encode visual and motor information is similarly 31 selective in the somatosensory domain. We recorded from 1423 neurons within the PPC of a human 32 clinical trial participant during objective touch presentation and during tactile imagery. Neurons encoded 33 experienced touch with bilateral receptive fields, organized by body part, and covered all tested regions. 34 Tactile imagery evoked body part specific responses that shared a neural substrate with experienced 35 touch. Our results are the first neuron level evidence of touch encoding in human PPC and its cognitive 36 engagement during tactile imagery which may reflect semantic processing, sensory anticipation, and 37 imagined touch. 38

39
The posterior parietal cortex (PPC) is critical to integrating sensory information into motor plans and 40 monitoring ongoing movement (1, 2). In recent studies, we identified at the level of single neurons, 41 evidence that the human PPC encodes a wealth of movement related information including movement 42 plans and trajectories, as expected, but also other variables such as cognitive motor imagery, action 43 semantics, observed actions, and memory (3-7). This richness of representation is made possible through 44 a partially mixed encoding in which single neurons represent multiple variables, allowing a relatively 45 small neuronal population (recorded through a small 4x4 mmm implanted microelectrode array) to 46 provide many movement related signals (6). Here, we explore whether this neuronal population also 47 encodes the somatosensory domain, given its often intimate association with movement planning, 48 initiation, and execution. 49 Multiple lines of evidence support somatosensory representations within the PPC. In non-human primates 50 (NHPs), neurophysiological studies demonstrate somatosensory processing in cell populations around the 51 intraparietal sulcus (IPS), where they are thought to play a role in higher-level cognition and perception 52 (8)(9)(10)(11)(12)(13)(14)(15)(16). Examples include monitoring of limb configuration (through convergent visual and proprioceptive 53 information) and sensing the space around the body (peripersonal space; convergent visual and tactile 54 information) (8-11). In humans, lesion and neuroimaging studies support similar representations (17-19). 55 Moreover, functional neuroimaging studies in humans demonstrate that experienced, observed, and 56 imagined touch activate overlapping regions of the PPC, suggesting its role in a multisensory, cognitive 57 processing of touch (20-25). While a sizeable body of literature has developed around somatosensory 58 representations within the PPC, several fundamental questions remain. At the single neuron level, how 59 are receptive fields to touch encoded? If bilateral information is represented, are the two sides 60 discriminable? To what extent are cognitive touch representations activated during imagined touch 61 encoded within the same neuronal populations? Does activity evoked during tactile imagery share a 62 neural substrate with experienced touch? 63 In a unique opportunity, we investigated touch processing at the level of single neurons in a tetraplegic 64 human subject recorded with an electrode array implanted in the left PPC for an ongoing brain machine 65 interface (BMI) clinical trial. In previous work, we have referred to the implant area as the anterior 66 intraparietal cortex, a region functionally defined in NHPs (3-6, 26). Here we will refer to the recording 67 site as the postcentral-intraparietal area (PC-IP), acknowledging that further work is necessary to 68 definitively characterize homologies between human and NHP anatomy. We recorded from a total of 69 1423 single neurons during the presentation of objective touch and during imagined touch to sensate 70 dermatomes above the level of the participant's injury. We found that human PC-IP neurons encoded 71 experienced touch at short latency (~100 ms) with bilateral receptive fields, covering all tested, sensate 72 regions within the head, face, neck, and shoulders. Tactile imagery evoked body part specific responses 73 that shared a neural substrate with experienced touch. Our results demonstrate for the first time, a high-74 fidelity, reproducible encoding of touch that can partially be reactivated during tactile imagery in a body 75 part specific manner. The latter represents a novel finding, thus far untestable in NHP models, and 76 suggests PPC involvement in the cognitive processing of touch. 77

79
We recorded from 1423 well isolated and multi-unit neurons in the PC-IP (left-hemisphere) of a high 80 cervical (level three to four; C3/4) spinal cord injured, tetraplegic human participant over 14 recording 81 sessions at approximately one-week intervals (on average, 101.64 ± 7.22 neurons recorded 82 simultaneously). Recordings were split across four tasks, designed to probe basic properties of the 83 neuronal population relating to both experienced and imagined touch. 84 85 86

PC-IP neurons encode bilateral tactile receptive fields 87
We first examined the hypothesis that PC-IP neurons encode tactile receptive fields to dermatomes above 88 the level of the participant's spinal cord injury (SCI). Tactile stimuli were delivered as rubbing motions at  89 approximately 1Hz, for 3 seconds. The subject was asked to keep her eyes closed to eliminate neural 90 responses arising from visual input. Tactile stimuli were presented to bilateral axial (forehead, vertex,  91 cheek, neck, back) and truncal (shoulder) body parts to determine the extent of body coverage of any 92 tactile representations among PC-IP neurons. As controls, touch was also presented to the bilateral hands 93 (insensate regions below the level of SCI), and a null condition was included (with no stimulus 94 delivered), to verify that touch related neural responses observed were not random. 95 A significant fraction of the neuronal population encoded touch to each of the tested body parts with 96 preserved somatosensation (Figure 1A, p<0.05, with false discovery rate (FDR) correction). These results 97 are consistent with bilateral encoding as the sensory fields included both left and right sides of the body. 98 No significant modulation was seen in response to stimuli delivered to the hands, or in the null condition. 99 Across the population, contralateral stimulation can be better discriminated (through a discriminability 100 index, DI) from baseline activity (Figure 1B, p<0.05, sign test). Many neurons demonstrated an exclusive 101 activation for touch to a single body part, although a substantial fraction of the population also 102 demonstrated activation to multiple body parts (shown as a bar plot in Figure 1C). Sample neuronal 103 responses of touch to various sites are shown in Figure 1D. A neuron that encodes touch to a particular body-part could 1) be entirely specific to that body part only, 126 2) may respond to others, but show proclivity for touch to an alternate, ipsilateral field, 3) may prefer the 127 corresponding, contralateral receptive field, or, 4) may show random patterns of activation to other 128 sensory fields. We performed a correlation analysis of the population responses across each stimulation 129 condition. The results are shown in Figure 1E and Figure 1F. The population responses to stimulation 130 on the right and left sides of the body were highly correlated ( Figure 1E), indicating that the pattern of 131 activation across the neuronal population for a body segment is largely equivalent across the left and right 132 sides of the body. For example, the population correlation between the left and right neck is similar to the 133 cross-validated correlation of the left neck to itself. We note that response patterns to non-identical body 134 parts (on either body side) are non-zero, suggesting a shared response to the simple presence of a stimulus 135 (or potentially the precise type of stimulus, e.g. rubbing motions), independent of the precise location of 136 the stimulus. A direct comparison of population correlation between matching body parts of the right and 137 left side with population correlations within a body side is shown in Figure 1F. 138 We found that population neural activity can be used to accurately classify body parts above the level of 139 injury, including differentiation of the body side ( Figure 1G). However, in line with the correlation 140 analysis, incorrect classifications tended to be for the matching body part on the opposite body side. 141 Tactile stimuli to insensate hand regions was frequently confused with the null condition, consistent with 142 the lack of any meaningful neural selectivity for these control conditions. 143 We examined bilateral coding at the level of individual neurons. We first investigated whether neurons 144 selective for body parts on the contralateral side were also selective for body parts on the ipsilateral side. 145 For each neuron we computed a linear model that described firing rate as a function of response to the 146 stimulated body part, independently for the contralateral and ipsilateral body sides. For each linear model, 147 we computed a cross-validated coefficient of determination (R 2 within ) to measure the strength of neuronal 148 selectivity for each body side. The R 2 within values for the left and right for each neuron are plotted against 149 each other in a scatter plot in Figure 2A. Most points cluster around the identity line, indicating that units 150 highly selective for body parts on the left were also highly selective for the right side. This bilateralism is 151 also reflected in the specificity plot shown in Figure 2B, in which it is evident that for most units the 152 strength of selectivity was comparable across the ipsilateral and contralateral sides, with a slight bias for 153 the contralateral side (p=0.04, sign test). This bias is consistent with the greater discriminability for touch 154 to the contralateral body than to the ipsilateral body seen in Figure 1B. 155 Next, we asked whether the precise patterns of responses observed for stimulation of the left body parts 156 generalized to the right side, and vice versa. In other words, how does the population level similarity in 157 coding for bilateral body parts manifest at the single unit level? To address this question, for each neuron, 158 we trained a linear model to predict firing rate as a function of stimulated body part using contralateral 159 data and predicted the firing rate of the ipsilateral data (and vice versa). The ability to predict across body 160 side was quantified as the R 2 across and compared to the R 2 within computed above ( Figure 2C and Figure  161 2E). We found that R 2 across and R 2 within clustered around the identity line, indicating a high similarity in 162 encoding between the two sides for corresponding receptive fields. Specificity plots are shown in Figure  163 2D and Figure 2F. The  index (see Methods) of each neuron, based on its response to touch for the right and 177 for the left body sides (data from Figure 2A) We explored PC-IP population encoding and single unit response latencies to tactile stimulation on the 188 contralateral and ipsilateral body sides. In a variation of the basic task paradigm, we used a capacitive 189 touch sensing probe to acquire precise measurements of the latency in neuronal response from the time of 190 tactile stimulation. We probed latency on the bilateral cheeks and shoulders. Again, as a control, we 191 included both hands in the task design. We compared latencies between the two sides at both the level of 192 the PC-IP neuronal population as well as at the single unit level (as detailed in Methods). 193 At the population level, we measured encoding latency as the time at which stimulated body parts were 194 discriminable based on a sliding window classification analysis. Encoding latency was short for both 195 body sides and was slightly shorter for contralateral (right) receptive fields (96 ms) than for ipsilateral 196 (left) receptive fields (104 ms) although this difference was not statistically significant. Figure  Neuronal receptive fields to tactile stimuli on the cheek are spatially localized 215 We explored the receptive field structure within a body part at finer spatial precision to begin to 216 characterize the sizes and shapes of receptive fields. We used a paintbrush to stimulate each of nine points 217 equidistantly spaced along the participant's cheek and neck, two centimeters apart, as shown in Figure  218 4A. A significant fraction of the neuronal population responded to tactile stimulation at each point, 219 although the fractional responsiveness of the PC-IP population appears greater to stimulation points on 220 the cheek than on the neck ( Figure 4B). The strength of modulation from baseline, measured by the 221 discriminability index, demonstrated a similar trend ( Figure 4C). These findings are consistent with the 222 results of Figure 1A, although all values are somewhat depressed, likely due to the gentler and spatially 223 localized sensory stimuli. At the population level, responses demonstrate spatially structured receptive 224 fields with stimulation of neighboring locations eliciting more similar activity than stimulation of distance 225 locations ( Figure 4D). Sample neuronal responses showing response as a function of stimulation site are 226 shown in Figure 4E. Most neurons preferred a single stimulation site and demonstrated progressively less 227 activity with increasing distance from that site. 228 To estimate the average size of neuronal receptive fields (see Methods for details), we first identified all 229 neurons demonstrating significant differential spatial representation of touch. For these units, categorized 230 by their site of preferred (peak) response, we fit the standard deviation of a Gaussian function centered on 231 the peak response to estimate the tuning width. Figure 4F shows the averaged responses for each 232 stimulation site. The full width at half maximum (FWHM) of the neuronal receptive fields ranged from site. The y-axis is a standard (z) score, representing how many standard deviations 255 the mean spiking activity for neurons at each stimulation site was from the mean 256 activity for that group of neurons at all sites together. 257 258

Tactile imagery evokes body part specific responses congruent with objective touch 259
Is the PC-IP recruited during tactile imagery? And if so, how might evoked neural responses compare to 260 those arising from objective touch? To address these questions, we analyzed population activity elicited 261 during a cue-delay-go tactile imagery task and compared the neural activity to that from objective touch 262 to matching body parts recorded during interleaved trials. For imagery, the participant was instructed to 263 imagine touch to the right (contralateral) cheek, shoulder, or hand with the same qualities as the objective 264 touch stimuli the participant experienced during interleaved trials. A null condition was included as a 265 baseline to measure neural activity without an objective or an imagined touch. 266 As with findings for objective touch, neuronal responses elicited by tactile imagery following the go cue 267 (during the imagery phase or epoch) were discriminably encoded ( Figure 5A, cross-validated 268 accuracy=92%). When the null condition was excluded from this analysis, the prediction accuracy was 269 higher, approaching 98%. High decode accuracy is consistent with the participant's compliance with task 270 instructions and implies that tactile imagery can elicit selective neural responses. 271 Consistent with previous results, a significant fraction of neurons encoded objective touch to the cheeks 272 and shoulders but not to the hands. In comparison, a smaller fraction of the neuronal population was 273 responsive to the cheek and shoulder during imagery of tactile stimuli ( Figure 5B). Of note, a significant 274 number of neurons responded to imagined touch to the hand, despite the hand being clinically insensate in 275 the study participant (and despite objective touch to the hand not eliciting neuronal activation). 276 Discriminability from baseline neural activity, measured by the discriminability index, demonstrated a 277 similar trend to the single unit responsiveness profile ( Figure 5C). 278 We used the population correlation measure to compare population level neural activity across conditions 279 ( Figure 5D). Neural activity during tactile imagery shared a neural substrate with responses evoked by 280 objective touch: representations for imagined touch and for experienced touch were more similar for 281 matching body parts fields than for mismatched body parts ( Figure 5E, permutation shuffle test p<0.05). The analyses above were restricted to the mean neuronal activity following the go cue (e.g. during 299 objective touch or during imagery) to allow a direct comparison with results reported for the previous 300 paradigms. We now expand this analysis. During the tactile imagery task, the participant heard a verbal 301 cue specifying a body part (verbal cue = "cheek," "hand," or "shoulder") followed approximately 1.5 302 seconds later by a beep instructing the participant to imagine the stimulus at the cued body part on the 303 right side of the body. This cue-delay paradigm is standard in the motor physiology literature and is used 304 to dissociate planning from motor execution related neural activity (3, 27-29). In our case, the cue-delay 305 was unique to the tactile imagery condition. We utilized the cue-delay task to begin to dissociate in time 306 whether neural activity during tactile imagery is consistent with the neural correlate of imagined touch. 307 To leverage the benefits of the cue-delay paradigm, we performed a dynamic classification analysis 308 (500ms windows, stepped at 100ms, see Methods). Results are shown as a matrix (see Figure 6). The 309 diagonal elements represent the cross-validated prediction accuracy for a specific time window. The off-310 diagonal elements represent how well the classifier generalizes to alternate time windows. Each row can 311 be interpreted as quantifying how well decision boundaries established for the diagonal time windows 312 generalize to other time windows. This analysis allows us to measure when the neuronal population 313 represents the different body parts (the diagonal) and whether population coding is similar or distinct 314 during the task phases (the off-diagonal). We are interested in two main phases of the task, the early 315 portion comprised of the cue and delay (cue-delay), and the later portion when the participant is actively 316 imagining the stimulus (go/imagery). Figure 6A schematically illustrates examples of possible results.

317
The examples are meant to be illustrative but are not an exhaustive list of possibilities. The population 318 may be selective exclusively during the imagery phase, during the cue-delay and imagery phases but with 319 distinct population coding, during the cue-delay and imagery phases with identical coding, or during the 320 cue-delay and imagery phases with partially shared and partially distinct coding. Each pattern would 321 suggest a different interpretation of various forms of cognitive processing that may be involved in tactile 322 imagery (see Discussion). 323 The results of our classification analysis ( Figure 6B) are most consistent with body part (or sensory field) 324 selectivity during both the cue-delay and imagery phases with partially shared and partially distinct 325 populations coding of the body parts between phases. The shared component is evident in the significant 326 generalization accuracy in the off-diagonal elements, a representative row of which is shown in Figure  327 6C (blue portion) where cross-validated accuracy generalizes from approximately 70% within the cue-328 delay phase to approximately 60% during the imagery phase. The distinct population activity between 329 phases is best revealed by a cross-validated Mahalanobis distance as it provides a sensitive measure of 330 change which is masked by the discretization process of classification (see Methods). The findings 331 demonstrate a significant change between the activity patterns in the cue-delay and the imagery epochs 332 (Figure 6C, black). each recorded unit. This resulted in the same matrices described above, but now each matrix represents 356 how information coding evolves for a single unit. Two complimentary analyses were then performed. In 357 the first, a cluster analysis was performed on the resulting matrices ( Figure 7A). Three clusters were 358 identified that most parsimoniously accounted for observed activity patterns (Bayesian information  359 criteria test for optimal number of clusters). Clusters roughly corresponded to temporal profiles with 360 selectivity during the cue-delay and imagery phases with similar coding (30%), and units exclusive to the 361 imagery epoch (33%) or the cue-delay epoch (37%). In the second analysis, time resolved classification 362 data were analyzed using a principle components analysis (PCA), the first three principle components of 363 which are shown in Figure 7B. Cognitive processing during the cue-delay and imagery epochs of the tactile imagery task shares a 378 neural substrate with that for objective touch 379 Finally, we look at how encoding patterns through time generalize between the tactile imagery and 380 objective touch conditions. The dynamic classification analysis above was applied both within the 381 imagery condition and across condition types (e.g., from imagined to experienced touch on the right side 382 and vice versa; Figure 8A). We found significant generalization when training the classifier on either the 383 cue-delay or the imagery phases of the imagery task and applying the classifier to the stimulus phase of 384 the objective touch condition (Figure 8A, lower-left panel). This implies that cognitive processing prior 385 to active imagery as well as during imagery share a neural substrate with experienced touch. A 386 visualization of significant generalization from imagery to experienced touch for the boxed region of 387 Figure 8A is shown graphically in Figure 8B. We did find an asymmetry in across-condition 388 classification; training a classifier on experienced touch did not generalize well to the imagined touch 389 condition ( Figure 8A, upper-right panel). This asymmetry is likely a consequence of the analysis 390 technique and may not be of physiological significance. Figure 8C illustrates the likely driver of the 391 asymmetry using classifiers trained on the first principal component of the population response. The 392 figure demonstrates how a decision boundary for relatively low signal-to-noise ratio (SNR) conditions 393 will generalize to a higher SNR class, but not vice versa. Sample neuronal responses that help to 394 understand single unit and population behavior are shown in Figure 8D.

419
We have previously reported that human PPC encodes many action variables in a high-dimensional and 420 partially mixed representation (5-7). This architecture allows many parameters to be encoded by a small 421 number of neurons, while still enabling meaningful relationships between variables to be preserved. Here 422 we show that neurons recorded from the same electrode array in the same clinical trial participant are also 423 selective for bilateral touch at short latency. Responses to objective touch are organized around body part, 424 sharing population representations between the left and right side. Additionally, tactile imagery elicits 425 body part specific responses that share a neural substrate with that for experienced touch. Furthermore, 426 we found neural selectivity during active imagery as well as during the cue and delay epochs that precede 427 imagery. The distinguishable population activity during these different phases indicates an encoding of 428 multiple cognitive processes that may include semantic association, memory, sensory anticipation, or 429 imagery per se. 430

431
Experienced touch representation in human PC-IP is bilateral, body part centric and at short 432 latency 433 Cortical processing of somatosensory information begins in the anterior portion of the parietal cortex 434 (APC) within four cyto-architectonically defined areas termed BA 3a, 3b, 1 and 2 (30-32). Each of these 435 four sub-regions represents primarily contralateral somatosensory information (33-40

Tactile imagery dynamically invokes multiple cognitive processes in human PC-IP 466
Functional neuroimaging data in humans suggest that a network of brain regions is activated during both 467 experienced and imagined touch. Involved brain regions include the PPC as well as portions of the insula 468 (in particular, the posterior insula), the amygdala, and the bilateral temporal cortices (23, 60, 61). Human 469 neuroimaging studies also support a role for the PPC in the interpretation of observed (not experienced) 470 touch to others (20-22, 25, 62, 63). These observations collectively suggest that the PPC represents a node 471 within a network of brain regions that support the shared processing of experienced, observed, and 472 imagined touch. 473 In motor neurophysiology, neural activity related to planning and execution are dissociated in time by 474 introducing a delay between the cue instructing movement, and the movement in response to the cue (27, 475 28). We have found that such distinctions between planning and execution are preserved during motor 476 imagery experimental paradigms in tetraplegic individuals (3). Here, a similar paradigm allowed temporal 477 dissociation in cognitive processing during tactile imagery. Single units demonstrated three dominant 478 response profiles: 1) a shared selectivity pattern between the cue-delay and imagery epochs, consistent 479 with cognitive engagement during all phases of the imagery task, 2) selectivity exclusively during the 480 cue-delay epoch but not the imagery epoch, and 3) selectivity exclusively during the imagery epoch but 481 not the cue-delay epoch. In a previous study, we found similarly heterogeneous responses during the cue, 482 delay and imagery epochs for imagined hand grasp shapes (64). These single unit temporal selectivity 483 profiles provide a basis for the population level findings of generalization in classification results between 484 the cue-delay and the imagery epochs ( Figure 6B and Figure 6C) but also a separation in neural state-485 space between these epochs ( Figure 6C). 486 We acknowledge a limitation: while our task is well designed to identify dynamic engagement of multiple 487 cognitive processes during tactile imagery, it is not adequate to precisely define the cognitive correlates of 488 observed neural activity. This will be a subject of future investigation. A conjecture is that neural activity 489 during the cue-delay and imagery epochs may reflect a combination of semantic processing of the verbal 490 cue, sensory memory, sensory anticipation of a tactile stimulus, and imagery itself (4, 7, 65). An 491 involvement of semantic processing is especially likely as we recently reported processing of read action 492 verbs within the same neuronal population (7). The current findings would extend these results to the 493 tactile domain and demonstrate neuronal selectivity for auditory cues (in addition to written text used in 494 the previous study). It may be that the neurons involved in semantic processing within the PC-IP are 495 shared whether for sensory (tactile) processing or motor planning. However, future work is necessary to 496 better characterize the properties of these neurons as they relate to body part specificity, modality 497 (sensory or motor) dependency, and other properties. 498 One concern with the use of all imagery experiments is that participant compliance cannot be externally 499 validated. This raises the possibility that the participant is not performing the task or is performing the 500 task in an unexpected manner. We think this is unlikely for three reasons. First, the subject by the time of 501 this study was well versed in performing cue-delayed paradigms in the motor domain using both motor 502 imagery and overt movements (e.g. 6). In Zhang and Aflalo et al. 2017, the participant's performance of 503 overt movements was perfect: the participant both performed the correct cued action and performed the 504 action at the go cue (e.g. no movements began prior to the go cue as validated by measurements of 505 electromyogram activity). Second, our current pattern of results that includes stable and accurate (near 506 100%) body part specific encoding within the cue-delay and imagery epochs, with a shift between epochs, 507 is consistent with the participant performing the task as instructed. At a minimum, it is consistent with the 508 participant's performing two distinct cognitive operations during the two primary phases of the task with 509 remarkable trial to trial consistency. Third, evidence for a shared neural substrate between experienced 510 touch and the imagined touch conditions (discussed more below) indicates that selective responses during 511 the imagery task are related to tactile cognition. 512 513

Tactile imagery shares a neural substrate with experienced touch in human PC-IP 514
We found that experienced touch and the cognitive activity evoked by imagined touch shared a neural 515 substrate within the PC-IP. Imagined touch to the cheek, for example, was more similar in representation 516 to experienced touch to the cheek than to experienced touch to the shoulder, and vice versa. Interestingly, 517 the overlapping neural representations between experienced touch and imagery were not limited to the 518 stimulus phase (objective touch and imagery) itself, but also extended to the cue-delay phase of the 519 imagery task. This overlap is consistent with our recent findings for shared neural representations 520 between imagined and attempted actions, and for shared neural representations between observed actions 521 and action verbs and is in line with our findings of a partially mixed architecture within PC-IP (5, 6, 66). 522 These studies are all also consistent with views in which cognition recruits sensorimotor cortical regions 523 (67-71). As with all passive neural recording studies, ours cannot establish a causal role for these neurons 524 in tactile cognition. Understanding the unique contribution of PC-IP neurons within the larger network of 525 brain regions engaged in cognitive touch processing remains to be explored. Nonetheless, our current 526 results provide the first human single unit evidence of a shared neural substrate between tactile imagery 527 and experienced touch. 528 A substantial fraction of the neuronal population activated in response to imagined touch to the hand, 529 where no response to objective touch was seen (insensate in the study participant). This suggests that 530 despite the lack of peripheral input from the hand due to the participant's spinal cord injury, the brain 531 maintains an internal representation of tactile sensations (72). The findings that intracortical 532 microstimulation produces discernable tactile perceptions from insensate body regions adds definitive 533 evidence for a maintained representation of somatosensory sensations after deafferentation (73, 74). 534 These findings will prove useful for bidirectional neural prostheses. 535 536

PC-IP and plasticity following spinal injury 537
The extent to which the human PPC reorganizes following SCI is unknown. Lesion studies in NHPs 538 suggest that BA 3b and 3a, 1 and 2, show altered sensory maps following SCI, in a manner dependent on 539 thalamic input from the afferent sensory pathways such as the dorsal column-medial lemniscus system 540 (75). With mid-cervical lesions, for instance, there is an initial loss of BA 3b hand representations, and a 541 slight expansion in face representation at approximately two years (75,76). Although significant axonal 542 sprouting has been demonstrated to occur at the site of deafferentation in the spinal cord, with increased 543 projections to brainstem nuclei, the changes observed in the somatosensory cortex are significantly 544 smaller (75,76). Moreover, the reorganization in higher order somatosensory centers such as the 545 secondary somatosensory cortex is even more restricted than in BA 3b (76). Against this background, we 546 acknowledge that while additional work probing cortical reorganization following SCI is necessary to 547 fully understand its electrophysiological consequences, our results within the current report provide 548 insight into the maintenance of basic tactile processing within the human PC-IP, and PPC, after SCI. 549 550

551
Multiple lines of evidence indicate a critical role for the human PPC in the integration of convergent 552 multimodal sensory information to enable complex cognitive processing and motor control. To date, 553 however, its processing of somatosensory information at the single neuron level has remained 554 fundamentally unexplored. In the unique opportunity of a neuroprosthetic clinical trial, we examined the 555 neural encoding of real and imagined touch within the human PC-IP. We found that local populations of 556 PC-IP neurons within a 4x4 mm patch of cortex encode bilateral touch sensations to all tested sensory 557 fields above the level of the participant's injury at short latency. A significant fraction of PC-IP neurons 558 encoded imagined touch with matching sensory fields to experienced touch. The activity in the delay 559 period of the task, between cueing and imagining touch, may reflect cognitive processes including tactile 560 semantics, sensory anticipation, as well as active imagery. Together, our results provide the first single 561 unit evidence of touch processing within the human PC-IP and identify a putative substrate for the 562 encoding of cognitive representations of touch, thus far untested in animal models. 563

566
Study participant 567 568 The study participant, NS, is a 60-year-old tetraplegic female with a motor complete spinal cord injury 569 (SCI) at cervical level C3-4 that she sustained approximately ten years prior to this report. She has intact 570 motor and sensory function to the level of her bilateral deltoids. NS was implanted with two 96-channel 571 Neuroport recording was assumed to be independent and no assumptions were made about the same units being 612 recorded on more than one study session. To minimize noise and low-firing effects in our analyses, we 613 used as a selection criterion for units, a mean firing rate greater than 0.5 Hz and a signal to noise ratio 614 (SNR) greater than 0.5. We defined SNR for the waveform shapes as the difference between their mean 615 peak amplitude and the baseline amplitude, divided by the variability in the baseline. 616 For measurements of neural latency to stimulus response (please refer to the task descriptions below for 617 more information), a custom capacitive probe was used to record the exact time of tactile stimulation. 618 This probe was built using a Raspberry Pi 2B and Adafruit Capacitive Touch Hat (Adafruit product ID 619 2340). The digital output (a binary output for touch or no touch) was transmitted through a BNC cable 620 into the NSP at an analog signal sampling rate of 2 kHz. 621 622 Task procedures 623 We used several experimental paradigms to probe various features of experienced and imagined touch 624 representations in the PC-IP. In each paradigm, the participant was instructed to keep her eyes closed. The 625 basic task structure comprised three phases: Each trial began with the presentation of a cue to the 626 experimenter (or an auditory cue in the tactile imagery condition, see specific task description below), 1.5 627 seconds in duration, indicating the stimulus (for example, touch NS's left cheek). This was followed by a 628 brief delay, 1 second in duration. Then written text appeared on the screen to signal the experimenter to 629 present the instructed stimulus for 3 seconds (in the tactile imagery paradigm, a beep indicated the "go" 630 signal for the participant). Exact time intervals varied depending on task. Trials were pseudorandomly 631 interleaved; all conditions were necessarily required to be performed at least once before they were 632 repeated. In tasks in which both left and right body sides (ipsilateral and contralateral to the implant, 633 respectively) were tested, stimuli were delivered to one body side at a time. 634 635 Neural responsiveness to touch 636 This task variant explored neuronal responsiveness and selectivity to objective touch to body parts 637 (receptive fields) with preserved somatosensory input (above the level of SCI). Body parts tested included 638 the forehead, vertex of the head, left and right back of the head, left and right cheeks, left and right sides 639 of the neck, and the dorsal surfaces of the left and right shoulders. As controls, the left and right hands 640 (clinically insensate) and a null condition (no stimulus presentation) were also included. Objective touch 641 stimuli were presented to each body part as finger rubs by the experimenter at approximately one per 642 second. Stimuli to the left and right body sides were delivered on separate trials to evaluate each side 643 independently. To ensure that any neural activity observed truly arose from experienced touch and not 644 from observed touch or other stimuli, NS was instructed to close her eyes throughout the task. She 645 additionally wore ear plugs to block auditory input. This task was performed on four separate days. On 646 each day, ten trials per condition were conducted. In total, we recorded from 398 sorted units on four 647 separate testing days. 648 649 Neural response latency 650 The purpose of this task was to determine the latency of neural response to objective touch for the left and 651 right sides of the body. Tested regions included the left and right cheeks, the left and right shoulders, and 652 as controls, the left and right hands (insensate). Objective touch stimuli were presented as in the task 653 above. Instead of finger rubs, however, a capacitive touch probe was used to enable precise delineation of 654 the actual time of contact (touch) before the onset of a neural response. This task was performed on eight 655 separate days, with eight trials per condition in each run of the task. In total, we recorded from 838 sorted 656 units. 657

658
Receptive field size 659 This task aimed to estimate the size of neuronal receptive fields to objective touch. Neural responses to 660 nine equally spaced points were evaluated, two centimeters apart, along a straight line from NS's right 661 cheek to her neck ( Figure 4A). Only the right side (contralateral) was tested in this task. The first of these 662 nine points was on the cheek bone or the malar eminence, and the ninth point was on the neck as shown. 663 In addition to the nine points, a null condition (no stimulus presentation) was also added. Stimuli were 664 presented through a paintbrush gently brushed against each of the points, at a frequency of one brush per 665 second. The paintbrush was employed to deliver spatially localized sensations without accompanying skin 666 distortion that could mechanically stimulate nearby sensory fields. Data were recorded on six separate 667 days. On each day, ten trials of each condition were tested. In total, we recorded from 585 sorted units. 668 669 Engagement during tactile imagery 670 This task was intended to establish whether PC-IP neurons are engaged by tactile imagery, and whether 671 neural patterns evoked by cognitive processing of imagined touch and experienced touch share a common 672 neural substrate (e.g. activate the same population of neurons in similar ways). In this variant, the 673 participant was presented with either objective touch stimuli or instructed to imagine the sensation of 674 being touched. NS was instructed to keep her eyes closed throughout. Objective stimuli were cued to the 675 experimenter with written words that appeared on the monitor. Because the participant's eyes were 676 closed, the participant did not receive any information about the body part that would be stimulated prior 677 to experiencing the touch. The cue was followed by a one second delay and then at the sound of a beep 678 (the "go" signal), rubs at 1Hz were presented with a metallic probe to either the left or right cheeks, 679 shoulders, or hands. During imagined touch trials, an auditory cue was presented to NS instructing her to 680 imagine being touched on her right cheek, shoulder, or hand. The auditory cue consisted of a voice 681 recording of the words "cheek", "hand", or "shoulder" with cue duration of approximately 0.5 seconds. 682 After a one second delay, at the sound of the beep, NS imagined touch to the cued body part. We asked 683 the participant to imagine the sensations as alternating 1Hz rubbing motions similar to what she 684 experienced during objective touch trials. A null condition (without objective or imagined touch), not 685 preceded by an auditory cue was used to establish a baseline neural response. Data were recorded on eight 686 separate days. Eight trials of each condition were performed on each testing day. In total, we recorded 687 from 838 sorted units. 688

690
In the analysis of data from the various task paradigms used in this study, we utilized several statistical 691 methods. Some were specific for certain tasks, but others were applicable to multiple sets of data from the 692 different paradigms. For ease of reference, we have described all methods together in this section. Where 693 necessary, we provide specific examples from tasks to help illustrate their use in our analysis. 694 695 Linear analysis 696 In order to determine whether a neuron was tuned (i.e., modulated by a specific condition), we fit a linear 697 model to its firing rate during both the stimulus presentation phase and a baseline time window. The 698 neuronal response during the stimulation phase window was summarized as the mean firing rate 699 computed between 0.5 and 2.5 seconds after stimulus presentation onset. The starting time of 0.5 seconds 700 was chosen to minimize the influence of variable experimenter delay in presenting the stimulus. The 701 baseline response was summarized as the mean firing rate during the 1. that index is of the same condition type, and 0 if the data point is of a different condition type. All 710 baseline samples were also assigned a 0, effectively pooling together baseline data independent of 711 condition. A unit is considered tuned to a condition if the t-statistic for the beta coefficient associated with 712 the condition is significant (p<0.05, false discovery rate (FDR) corrected for multiple comparisons). 713 714 Discriminability index (DI) 715 We wished to derive a measure that quantifies how well neural activity can be discriminated from 716 baseline (e.g., pre-stimulus) activity. In other words, we wanted to capture how "well" or how "strongly" 717 a specific sensation (experienced or imagined) is encoded. We developed a cross-validated 718 discriminability index. As with the linear analysis described above, neuronal activity was summarized as 719 the mean firing rate during the stimulation phase window, defined as 0.5 to 2.5 seconds after the onset of 720 stimulus presentation. Baseline phase activity was summarized as the mean firing rate during the 1.5 721 second window before the stimulus onset presentation cue. We used correlation to compare the population neural representations of various tested conditions 731 (stimulus presentations) against each other in a pairwise fashion. Correlation was chosen over alternative 732 distance metrics (such as Mahlanobis or Euclidean distance) because it provides an intuitive metric of 733 similarity that is robust to gross changes in baseline neural activity across the entire neural population. 734 Results were qualitatively similar for alternate distance measures (specifically Mahlanobis distance). In 735 performing correlation analyses, we quantified the neural representations as a vector of firing rates, one 736 vector for each condition with each vector element summarizing the response of an individual unit. As 737 before, neural activity was summarized as the mean firing rate during the stimulation phase window, 738 defined as 0.5 to 2.5 seconds after the onset of stimulus presentation. To test whether the difference between any pair of conditions was statistically significant, we used a 748 shuffle test applied to the correlations computed over the 250 random splits. To illustrate, in Figure 5E  749 we applied this analysis to test whether the correlation between experienced and imagined cheek touch 750 was significantly different from that of experienced cheek touch and imagined shoulder touch. The true 751 difference in the correlations was computed as the difference in the mean correlations between 752 experienced and imagined cheek touches (over the 250 splits) and the mean of the correlations between 753 experienced cheek touch and imagined shoulder touches. We then randomly shuffled the two distributions 754 together (2000 times) and computed the difference in the mean correlations for each shuffle. The 755 distribution of shuffled differences served as the null distribution, against which we compared the true 756 difference to determine significance. 757 758

759
Classification was performed using linear discriminant analysis with the following assumptions: one, the 760 prior probability across tested task epochs was uniform; two, the conditional probability distribution of 761 each unit on any epoch was normal; three, only the mean firing rates differ for unit activity during each 762 epoch (covariance of the normal distributions are the same for each); four, firing rates for each input are 763 independent (covariance of the normal distribution is diagonal). The classifier took as input a matrix of 764 average firing rates for all sorted units. The analysis was not limited to significantly modulated units to 765 avoid "peeking" effects. Classification performance is reported as prediction accuracy of a stratified 766 leave-one-out cross-validation analysis. The analysis was performed independently for each recording 767 session and results were then averaged across days. 768 769

770
The purpose of this analysis was to 1) assess the degree to which tactile information is bilaterally 771 encoded, and to 2) assess the generalizability or similarity in representation for each side to the other (i.e., 772 whether the right and left sides of the body are coded in a similar manner). Only units demonstrating 773 selectivity, that is, differential coding for at least one segment of the body were included in this analysis. 774 To address the former, for each neuron, we computed the cross-validated coefficient of determination 775 (R 2 within ) to measure how well a neuron's firing rate could be explained by the responses to the sensory 776 fields. The R 2 within metric was computed separately for responses to the left (ipsilateral) side and the right 777 (contralateral) side of the body and compared to determine whether the population encoded 778 representations for one body side more robustly than the other. 779 To address the latter, for each neuron, we computed a regression model using neural data from the 780 ipsilateral side of the body and predicted neural responses for the contralateral side of the body (and vice 781 versa). Predicted responses were compared against true responses to compute a generalization R 2 across 782 metric. This generalization R 2 across was then compared against the cross-validated metric (R 2 within ) to 783 determine how similar sensory fields were encoded across the left and right sides of the body at the single 784 unit level. 785 786 Neuronal specificity index 787 The general formula we used to evaluate the degree of specificity (or specificity index) was: 788 789 and ‫ݕ‬ values correspond to those shown in Figures 2A, 2C, and 2E 805 We quantified the neural response latency to touch stimuli at the level of the neural population. Prior to 806 the analysis, trials were aligned by touch onset as detected by the capacitive touch sensor (ground truth). 807 Latency was estimated using a sliding window decode analysis: decode performance was computed with 808 k-fold cross-validation of a linear classifier trained over a sliding window through time (linear 809 discriminant analysis with equal diagonal covariance matrices field as a function of its preferred point of stimulus delivery, we fit a Gaussian model to the average 826 responses grouped by the preference of the neuron (i.e., data in Figure 4F). This description relates to Figure 6B. We performed a sliding-window classification analysis to quantify 837 the strength and temporal dynamics of population coding in the tactile imagery task. In this task, the 838 participant heard an auditory cue specifying a body part ("cheek", "hand", or "shoulder") that lasted 839 approximately 0.5 seconds, followed by an approximately two second delay, and finally a beep 840 instructing the participant to initiate imagining a touch sensation at the cued body part. This task could 841 engage at least three cognitive processes: 1) semantic processing of the cue; 2) preparation/anticipation 842 for imagery; 3) imagined touch per se. We used a dynamic classification analysis to understand how the 843 neural population evolved through the course of the trial to determine whether the population was best 844 described as mediating a single cognitive processes or multiple cognitive processes. In brief, the analysis 845 consisted of creating a dataset that consisted of the population response measured in small temporal 846 windows throughout the course of the trial. We trained a classifier separately on each temporal window 847 and applied each classifier to both temporal windows. We believe that this technique, by helping us to understand when information appears and how 863 information compares across task phases, provides a valuable approach to understanding how population 864 activity relates to the underlying cognitive processes. For example, if neural decoding reaches 865 significance only after the go cue, neural activity would be inconsistent with semantic or anticipatory 866 processing. Alternatively, if neural processing begins with the cue, and the same pattern of neural activity 867 is maintained throughout the trial, with no changes during the active imagery phase, then the data would 868 be inconsistent with processing imagined touch per se. 869 870 The classification analysis described above was used to measure general similarity of the population 871 response to the tested conditions across time. However, to explicitly test whether population activity was 872 changing, we used Mahalanobis distance as our measure. This is necessary as classification involves a 873 discretization step that makes the technique relatively insensitive to changes in neural population activity 874 that do not cross decision thresholds. Mahalanobis distance, being a proper distance measure, is a more 875 sensitive measure of change. To illustrate, imagine that a classifier is trained on time point A and tested 876 on time point B. At time point A, the means of the two classes are 0 and 1 respectively and at time point 2 877 the means are 0 and 4 respectively. All classes are assumed to have equal but negligible variance (e.g. 878 0.01) in this example. When trained on time point A, the classifier finds a decision boundary at 0.5. with 879 100% classification accuracy. When tested on time point B, with the same 0.5 decision boundary, the 880 classifier again is 100%. Naively, this could be interpreted as signifying that no change in the underlying 881 data has occurred, even though the mean of the second distribution has shifted. 882 883 Separation in neural activity between the cue-delay epoch and the imagery epoch was quantified using a 884 cross-validated Mahalanobis distance computed between the observed neural activity at a time point and a 885 reference (baseline) defined as the neural activity immediately following the presentation of the auditory 886 cue, from .25 to .75 seconds. Distances were measured separately for each of the three conditions and 887 then averaged. The mean and standard error on the mean (SEM) were computed across sessions for the 888 cross-validated distance measures and plotted in Figure 6C. Activity during the cue-delay epoch and the 889 go epoch were compared using a rank-sum test of the averaged activity during the phase averaged 890 responses across sessions. 891 892 Temporal dynamics of single unit activity during tactile imagery task: within category 893 We wished to understand the behavior of single neurons that led to the temporal dynamics of the 894 population. The temporal dynamics of single unit activity during the imagery task (for the imagined touch 895 conditions only) were quantified using both a cluster analysis (Figure 7A), and a principle component 896 analysis (PCA, Figure 7B). For both, a sliding-window classification analysis was first performed on 897 each sorted unit from all testing days in the same manner as described above for the population activity, 898 with the exception that classifier took as input a vector of the firing rates for a single unit as opposed to a 899 matrix of the firing rates for all units recorded in a single session. This allowed a quantitative description 900 of the temporal dynamics for each sorted unit. We next restricted neurons to those whose 90 th percentile 901 accuracy was at least 50%. This was to ensure only neurons with some degree of significant selectivity 902 were used for the cluster analysis. Next, a cluster analysis was performed on these matrices using K-903 means clustering and the cosine distance metric (chosen to provide a measure of temporal similarity in 904 neural activity profiles, robust to the decode accuracy itself.) We tested cluster sizes from 2 to 5 clusters 905 and used Bayesian information criteria (BIC) to identify the most parsimonious number of clusters for the 906 observed data. In the second analysis, a principal component analysis was applied to the dynamic 907 classification matrices with individual neurons counting as the independent observations. PCA has 908 become a standard method for describing the behavior of neural populations. Typically, PCA is applied to 909 firing rate measurements of neurons. However, in our case, we were less interested in capturing the main 910 modes of variability with respect to individual conditions, but instead wanted to capture the main modes 911 of variability with respect to the temporal dynamics of information encoding. 912 913

914
This description pertains to Figure 8A. Time-resolved classification analysis was performed using linear 915 discriminate analysis with assumptions and cross-validation procedures as described for within category 916 decoding above. For this analysis, both the experienced touch condition and the imagined touch condition 917 within the tactile imagery task were used. For the experienced touch category, only stimuli to the right 918 cheek, shoulder and hand were used in the analysis; neural activity from the left was not used, to try and 919 match the conditions for the imagined touch conditions in which only right cheek, shoulder, and hand 920 were tested. Classifiers were trained within category and applied to either itself or the other category 921 during each fold. Predictions across folds of the cross-validation procedure were used to compute decode 922 accuracy. This enables us to understand how well the neural representation of the two categories 923 generalize to each other, as well as how well neural representations generalize from one epoch (cue-924 delay) to another (stimulus: imagined or experienced touch